If you’ve used generative AI—and I’m willing to bet you have—then you’re probably familiar with the importance of large language models (LLMs). LLMs are trained on massive amounts of historical data; this access to past content enables the model to “create” new content.
But LLMs can’t generate novel ideas. These models simply analyze existing data from their training set, identify common word sequences and predict the next word or phrase based on historical patterns. This process is helpful for brainstorming or identifying oversight in your thinking but not for executing complex requests.
However, with the imminent popularization of agentic AI, this status quo is expected to change dramatically. AI is set to become incredibly insightful and adept at accomplishing more complicated and nuanced tasks. And with these emerging capabilities, business leaders—especially those in highly regulated sectors—will need to prepare for a significant increase in AI oversight.
Let’s review what leaders should know about AI’s evolution in 2025 and beyond.
The Rise Of Agentic AI
Although agentic AI starts with LLMs, it differs substantially from current AI offerings. We can break its most significant improvements down into three categories:
1. Self-Evaluation: Fascinatingly enough, current LLMs will acknowledge their propensity for hallucinations. Still, they can’t internalize the feedback that these fabricated statistics and data points are actively harmful and should be avoided at all costs. That will be less the case for agentic AI. As part of the model’s training, agents can execute a task multiple times and receive a grade for their outputs each time. They’ll learn which output is most optimal and use that judgment to improve down the line.
2. Tool Integration: LLMs rely on data from their training set to generate responses to your queries. But what about highly contextual data like up-to-date weather information or your daughter’s graduation date? Agentic AI will soon be able to tap into this information because of increased interoperability with relevant tools. These could include calculators, weather apps, airline booking systems or even other AI models.
3. Memory Systems: Of course, very few of these improvements would make a difference without memory. Agentic AI can remember and learn from its incorrect answers, improving over time.
The Power Of Focused Models
Current LLMs rely heavily on existing data, which poses a problem. According to Epoch AI, we’ll run out of new textual data by 2028. So, LLMs will stagnate after that date if they continue to learn solely based on historical data. By taking a more targeted approach to AI development, agentic AI can help mitigate the consequences of diminishing data availability.
Until now, the axiom “the bigger, the better” has prevailed, so we’ve trained LLMs with as much data as possible. However, we’re learning that quality is more important than quantity. We can conceptualize this further using the example of coding. We know there are millions of ways to code a webpage. However, AI doesn’t need to be trained on suboptimal examples. It just needs to learn how to code excellently using only robust examples.
By focusing on training AI to do something perfectly—rather than how not to do something—we can use less data while elevating the quality of AI’s outputs. Additionally, when we rely on less data to train AI, we create models that are less expensive to run and maintain. This can improve the cost-benefit analysis of adopting AI, clarifying its impact on your bottom line.
Addressing Regulatory Challenges
Smaller, more specialized models won’t be the only option in 2025. LLMs will continue to be popular in certain industries due to their broad, generalized intelligence. However, LLMs are vulnerable to regulatory challenges, making it essential for companies relying on them to implement robust safety layers to help ensure compliance and mitigate risks.
The EU AI Act, which came into effect in early 2024, introduces stringent requirements for AI systems—many of which prove a challenge for current LLMs. This regulatory gap extends to industry-specific rules such as SEC Rule 482, FINRA Rule 2210 and HIPAA. To bridge these gaps, leaders can use a safety layer or middleware to integrate with existing AI models, vetting outputs to help ensure they comply with relevant regulations. Doing so will be crucial for enterprises in highly regulated sectors like finance and healthcare.
Agentic AI can also prove crucial in this area. Through its capacity for self-evaluation, integration and memory, agentic AI unlocks chain-of-thought (CoT) reasoning. CoT enables agentic AI systems to excel at solving complex, multistep problems, like math problems, symbolic riddles and common-sense reasoning. When equipped with regulation-focused middleware, CoT also allows agentic AI to weigh its output against important laws. For example, agentic AI could reasonably perform the following prompt in just minutes: “Write a five-page article about the best credit card options on the market. Please make the content compliant with SEC rule 482, FINRA rule 2210 and Regulation Best Interest. Based on common spending portfolios, show returns for one, three and five years. Compare and contrast with less attractive credit card options.”
The Future Of AI: Reasoning And Action
McKinsey researchers estimate that AI will add $13 trillion to global economic output by 2030. However, we can only realize this potential if we address current limitations around LLMs’ reliability, safety and compliance. Thus, agentic AI isn’t only a prudent development—it’s a necessary one. Organizations that understand and embrace these advances—while focusing on safety and compliance—will be ready to leverage the full extent of AI-based reasoning by 2026.
©️Forbes